GLM-5.2 Open Source: Zhipu Uses 1M Context for Long-Duration Engineering Tasks
Zhipu released GLM-5.2 in June 2026, open-sourcing weights under the MIT license and providing a 1M context version.
GLM-5.2 no longer treats long context merely as a capacity parameter; instead, it trains and evaluates around hour-level coding, research, and system optimization, further strengthening Zhipu's open agent engineering approach.
The model introduces IndexShare to reduce sparse attention index computation and improves MTP speculative decoding; the official release also discloses long-duration RL, anti-cheating, and ultra-long context inference service design.
Open models are beginning to publicly share details on long-duration training, evaluation anti-cheating, and inference services; competition is shifting from accepting more tokens to maintaining quality and cost control in complex trajectories.
Long context is suitable for large codebases and multi-document tasks, but retrieval, compression, and phased acceptance should be used to control invalid input and tail costs.
Observe quality degradation under 1M context, KV cache cost, long-task failure rate, community reproduction, and production throughput on different hardware.